I wrote the following paper for my first-year philosophy seminar Skepticism and Knowledge, which was taught by Professor Catherine Elgin.



In Meditations,1 Descartes argues that we cannot fully trust the information given to us by our senses, as they have misled us in the past. We will never be able to rule out the possibilities that we are dreaming, that we are a brain in a vat, or that the universe (and all of our memories) sprang into existence 15 minutes ago. I find Descartes’ argument compelling, if not completely convincing, and I know many other readers feel the same way. After reading his work, I grudgingly concede to the inevitability of universal skepticism, admitting that almost nothing I perceive about reality can be known for certain.

At the same time, I continue living a fairly conventional life. I fearlessly walk over bridges, believing them to be stable. I recycle plastic bottles, hoping to protect the environment and slow climate change. And I try to treat other people with respect, believing that they have a consciousness and emotions similar to mine. How is it that I can perform these actions with sufficient confidence in my underlying beliefs when I have simultaneously conceded to universal skepticism? In other words, how is it possible to have more confidence in a high-level idea (“climate change is real”) than an extremely fundamental idea (“I am not a brain in a vat”)?

This presents a serious epistemic problem for anyone who accepts Descartes’ argument. If left unanswered, we will either have to abandon all confidence in our everyday actions, or we will be forced to admit that our system of beliefs is inconsistent. At a first glance, this problem appears impossible to resolve. It seems that knowledge must be built from the bottom up as it is done in a mathematical proof: confidence in complicated, specific statements depends on confidence in more basic ones. Under this logical framework, a high-level statement such as “Earth’s average temperature has risen roughly 1.4 degrees Fahrenheit in the past century” (call this B) depends on confidence in numerous prerequisite statements. These range from other high-level statements like “atmospheric data collected in the past century is accurate” all the way down to “the Earth, and reality in general, exists as we perceive it and is not a simulation” (call this A). If we are uncomfortable with accepting A, this logical framework tells us that we cannot be confident about B. This presents a contradiction because I am confident about B (as are many other people).

Fortunately, I believe there is a way to resolve this inconsistency by introducing the notion of epistemic context, which I define to be the set of assumptions and prerequisite statements that implicitly accompany any statement of knowledge. In this paper I will attempt to motivate the idea of epistemic context, justify its existence and role in our model of knowledge, and show how it enables a rational thinker to have confidence in high-level statements like B while simultaneously doubting low-level statements like A.

The motivation for epistemic context comes from the observation that certain statements seem to exist on different levels of understanding. For example, consider the two statements N: “Napoleon was a bad emperor” and P: “there are an infinite number of prime numbers.” While both statements postulate some property of our universe, it seems clear that N comes from a different sphere of knowledge than P. To understand what N is claiming, you must have a thorough understanding of human history and some criteria for what makes a good or bad ruler. But all you need to understand P is the mathematical concepts of number and divisibility. The methods used to justify the two statements also differ. Justifying N would require an analysis of all of Napoleon’s actions, their long-term effects on global affairs, and perhaps an appeal to ethics. Justifying P would involve some basic number theory and a proof by contradiction. This difference suggests that we should place N and P into two different categories of knowledge, or epistemic levels, separated by the presence of distinct epistemic contexts.

It feels natural to think about knowledge in this layered manner, likely because this is the way our human psychology processes information. When trying to solve a math problem within the epistemic level of P, I never even consider employing facts about European history, and vice versa. My brain automatically restricts itself to the epistemic context in focus. A reader might object that I am taking too much of a human-centric view of knowledge. Perhaps by drawing inspiration from the compartmentalization of the human thought process, I am ignoring some more universal property of knowledge. However, I would respond that knowledge is in fact a uniquely human pursuit. After all, it is impossible to have knowledge of a statement without the use of a brain, or a similar thinking organ.

Now that I have motivated the idea of epistemic context, I will demonstrate why it must exist. In other words, I hope to show that any accurate model of knowledge2 must take epistemic context into account when deciding the truth of a statement. Consider the naive model of knowledge that omits all epistemic context from consideration. More precisely, we can think of the model as a function \(f\) that takes in two inputs: \(S\) and \(U\). The first input, \(S\), is a statement in the form of an English sentence (“Earth’s average temperature has risen roughly 1.4 degrees Fahrenheit in the past century”). The second input, \(U\), represents the current configuration of reality. For example, if it turns out our modern conception of physics is relatively correct, \(U\) would encode the exact position of every atom in the universe, the spin of every electron on every atom, and so on. On the other hand, if it turns out reality is a giant simulation, \(U\) could encode the exact layout and memory configuration of the supercomputer that our universe is being simulated on. In any case, in this naive model of knowledge we have the function \(f(S, U)\) output a Boolean value: whether or not \(S\) is true given the objective state of the universe \(U\).

What happens if we try to plug in the statement \(S\) = “the US economy has been doing well recently”? Problems with \(f\) can be seen immediately. First of all, human languages like English are imprecise and ambiguous. What is the exact definition of the word “recently”? We could agree to be rigorous and say that “recently” refers to the time period marked by the latest revolution of Earth around the sun (one year). But what about the definition of “economy”? This seems much harder to define precisely. And what about the nuanced (and a bit subjective) concept of an economy “doing well”? This alone suggests that such a function \(f\) cannot exist, as it seems impossible to give a consistent set of necessary and sufficient conditions for every concept in the English language. But even if we ignore the messiness of language and assume that \(f\) has access to a precise definition for every English word, we run into another problem. Imagine the scenario in which \(U\) reveals I am the unknowing star of a TV series called “The Gabe Wu Show,” in which I have been led to believe in fake countries, including the US. In this reality, there is no such thing as “the US economy” – it only exists in the show’s script. What would \(f(S, U)\) evaluate to? It can be neither true nor false, because either answer would acknowledge that the US economy is something that exists in the objective universe, which it is not. However, \(S\) is still a well-formed statement that we should be able to ask. These difficulties tell us that such a formulation of \(f\) cannot exist in the way we want it to, so a context-free model of knowledge cannot make sense.

I will now show that a more nuanced model of knowledge that takes epistemic context into account can remedy this issue. The main idea is that we use epistemic context as a prerequisite for returning a truth value. If the epistemic context is violated, then we are allowed to leave the truth value indeterminate. Formally, we have a function \(g(S, U, C)\) that takes in \(S\) and \(U\) as before, but also receives the epistemic context \(C\) surrounding \(S\). For example, if \(S\) is “the US economy has been doing well recently,” then \(C\) might consist of the set of assumptions “physical space exists as we perceive it”, “the US is a real country with an economy that has existed for over a year”, “there is a relatively objective metric for measuring the performance of an economy”, and so on, along with precise definitions of all of the concepts referenced in \(S\). Then we define \(g(S, U, C)\) to return a value of true or false as long as \(U\) satisfies all of the preconditions in \(C\). If \(U\) does not satisfy \(C\) (the “Gabe Wu Show” scenario, for example), then \(g\) is indeterminate. \(S\) only has relevance inside its given epistemic context. To use mathematical terminology, this can be thought of as restricting the domain of \(g\) to values of \(U\) that satisfy \(C\).

By relaxing our model of knowledge to a context-aware function \(g\), we avoid the problems we had with the context-free function \(f\). Instead of forcing a Boolean value onto every statement in every scenario, we use epistemic context as a “sanity check” to protect our function from judging statements that simply do not apply to a given universe. The main weakness of \(g\) is that it cannot be universally applied. Under \(g\), when I say “it is raining,” it is possible that my statement is neither true nor false (which is certainly disturbing to a classical mathematician who uses the law of the excluded middle). However, I believe this is a price we must pay: we sacrifice the aesthetic of universal truth for a more accurate and consistent model of knowledge.

Now that we have motivated and justified the existence of epistemic context, we can use it to resolve our original contradiction. Recall the problem raised at the start of this paper. How is it possible to have more confidence in a high-level statement like B than we do in a low-level statement A that acts as a prerequisite to B? The answer is that in the context-aware model of knowledge, we can define confidence in a statement X to be the probability with which you judge a statement to be true, conditional on its epistemic context Y being true. In other words, it is the probability that \(g(X, U, Y)\) is true, taken over all scenarios \(U\) in which \(g\) returns a value. Mathematically, confidence in a statement X with epistemic context Y is given by \(Pr(X | Y)\). By Baysian probability, we get:

$$\text{Confidence in } X = Pr(X | Y) = \frac{Pr(X \wedge Y)}{Pr(Y)}.$$

This value can be greater or less than the confidence in Y, depending on what the statements actually say. Applying this to the specific problem at hand, we see that it is possible to have more confidence in the presence of climate change (B) than we have in the nature of reality (A), even when A is a prerequisite included in the epistemic context of B. Notice that we never claim that confidence in B must be greater than A – for some people it probably isn’t! But for the average educated, rational human in 2021, including me, this is the case.3

The main takeaway from this paper is that our confidence in any statement of knowledge is quantified not by the probability of the intersection of the statement and its epistemic context, but by the conditional probability of the statement given its context. This realization enables us to accept Descartes’ argument for universal skepticism while still having confidence in high-level statements. This is a very positive result – it means that I can continue to justify my conventional lifestyle of walking over bridges and recycling plastic bottles, while also being a universal skeptic. The notion of epistemic context served as an invaluable tool for arriving at this conclusion. Further research on this concept would be very important for understanding the true nature of knowledge.



  1. Descartes, Rene. Meditations 1 and 2. https://www.earlymoderntexts.com/assets/pdfs/descartes1641.pdf ↩

  2. I will use the term “model of knowledge” to refer to the underlying framework in which we evaluate the truth of any statement. In this paper, I compare two models of knowledge: one that omits epistemic context, and one that includes epistemic context. I hope to demonstrate that the latter is more accurate than the former. ↩

  3. The exact probabilities are of course extremely hard to gauge, but I would estimate that I have at least 95% confidence in B (given its epistemic context), and perhaps 80% confidence in A. Justifying this last value would be extremely difficult because the epistemic context of A is so minimal. ↩